Chengdu Institute of Computer Applications, Chinese Academy of Sciences, Chengdu, China.
Chongqing Institute of Green and Intelligent Technology, Chinese Academy of Sciences, Chongqing, China.
BMC Med Inform Decis Mak. 2022 Feb 26;22(1):51. doi: 10.1186/s12911-022-01788-8.
In the big wave of artificial intelligence sweeping the world, machine learning has made great achievements in healthcare in the past few years, however, these methods are only based on correlation, not causation. The particularities of the healthcare determines that the research method must comply with the causality norm, otherwise the wrong intervention measures may bring the patients a lifetime of misfortune.
We propose a two-stage prediction method (instance feature selection prediction and causal effect analysis) for instance disease prediction. Feature selection is based on the counterfactual and uses the reinforcement learning framework to design an interpretable qualitative instance feature selection prediction. The model is composed of three neural networks (counterfactual prediction network, fact prediction network and counterfactual feature selection network), and the actor-critical method is used to train the network. Then we take the counterfactual prediction network as a structured causal model and improve the neural network attribution algorithm based on gradient integration to quantitatively calculate the causal effect of selection features on the output results.
The results of our experiments on synthetic data, open source data and real medical data show that our proposed method can provide qualitative and quantitative causal explanations for the model while giving prediction results.
The experimental results demonstrate that causality can further explore more essential relationships between variables and the prediction method based on causal feature selection and effect analysis can build a more reliable disease prediction model.
在席卷全球的人工智能浪潮中，机器学习在过去几年中在医疗保健领域取得了巨大成就，但这些方法仅基于相关性，而不是因果关系。医疗保健的特殊性决定了研究方法必须符合因果规范，否则错误的干预措施可能会给患者带来终身不幸。
我们提出了一种用于实例疾病预测的两阶段预测方法（实例特征选择预测和因果效应分析）。特征选择基于反事实，并使用强化学习框架设计可解释的定性实例特征选择预测。该模型由三个神经网络（反事实预测网络、事实预测网络和反事实特征选择网络）组成，并使用演员-批评者方法对网络进行训练。然后，我们将反事实预测网络作为一个结构化因果模型，并改进基于梯度积分的神经网络归因算法，以定量计算选择特征对输出结果的因果效应。
我们在合成数据、开源数据和真实医疗数据上的实验结果表明，我们提出的方法可以在提供预测结果的同时，为模型提供定性和定量的因果解释。
实验结果表明，因果关系可以进一步探索变量之间更本质的关系，基于因果特征选择和效应分析的预测方法可以构建更可靠的疾病预测模型。
